Technical WhitepaperTIPS Certified

Revenue Intelligence for Sports Facilities

Dynamic Pricing System for Perishable Inventory Optimization

Lab: BlueX ResearchAuthors: J. Hwan, S. YoonPublished: January 2026Status: TIPS Certified

Abstract

We present a comprehensive machine learning system for sports facility revenue optimization, achieving 98% prediction accuracy on demand forecasting across 16 metropolitan regions in South Korea. The system integrates three novel research modules: (1) a weather-demand correlation model using Korea Meteorological Administration (KMA) ASOS data with temporal lag analysis, (2) a LightGBM-based dynamic pricing engine with 19 engineered features constrained to ±60% price elasticity, and (3) a hybrid collaborative-content filtering recommendation system using matrix factorization with cosine similarity.

Our foundational insight treats sports facility reservations as perishable inventory— economically equivalent to airline seats and hotel rooms—yet managed with static pricing models that fail to capture time-decay value, weather elasticity, and demand-supply dynamics. This research introduces yield management techniques to an industry that has historically lacked them.

1

The Economics of Perishable Inventory

The Time-Decay Problem

A tennis court at 6:00 PM has economic value. At 6:01 PM, that specific time slot's value becomes permanently zero. Unlike physical goods that can be stored and sold later, time-based services have an absolute expiration—structurally identical to airline seats after departure.

1.1 Industry Comparison

Airlines and hotels solved the perishable inventory problem in the 1980s with revenue management systems. American Airlines' SABRE system alone generated $1.4B in incremental revenue through dynamic pricing. The core principle: some revenue is better than zero revenue.

IndustryYield ManagementPrice VariationWeather Factor
AirlinesSABRE, Amadeus200-400%Minimal
HotelsExpedia, Booking.com150-300%Seasonal
Ride-sharingSurge pricing100-500%High (rain)
Sports FacilitiesNone (before our research)0%Ignored

1.2 The Missing Elasticities

Our analysis identified two critical price elasticities absent from sports facility operations:

Time × Occupancy Elasticity

Price should vary based on proximity to time slot and current booking rate. An empty 7 PM slot at 6:30 PM should trigger automatic price reduction.

P(t) = P_base × (1 - decay(Δt)) × occupancy_factor

Weather Impact Elasticity

Outdoor court value correlates strongly with precipitation probability. 60%+ rain forecast reduces booking completion by 340%.

V(weather) = V_base × (1 - β × P(rain)) for outdoor

Revenue Optimization Philosophy

Our system optimizes for total revenue across all inventory, not per-slot margin. This leads to a counterintuitive principle:

"A quick sale at ₩20,000 > An empty slot at ₩30,000"

The ₩20,000 is captured revenue. The ₩30,000 is ₩0.

2

Module 1: Weather-Demand Correlation

We developed a weather-aware demand prediction system integrating Korea Meteorological Administration (KMA) data. Unlike simple "rain = bad" heuristics, our model captures temporal lag effects—booking behavior changes hours before actual weather events.

KMA ASOS Integration

Automated Surface Observing System

단기예보 (Short-term)

  • Forecast window3 days
  • Update frequency3 hours
  • Spatial resolution5km grid

중기예보 (Medium-term)

  • Forecast window10 days
  • Update frequency6 hours
  • Spatial resolutionCity-level

Coverage

16 metropolitan regions including Seoul, Busan, Incheon, Daegu, Daejeon, Gwangju, Ulsan, Sejong, and 8 provincial capitals.

SeoulBusanIncheonDaeguDaejeonGwangjuUlsanSejong+8 more

Key Statistical Findings

340%
Cancellation increase
when rain prob > 60%
-4h
Booking lag
behavior change precedes weather
78%
Transfer capture
outdoor → indoor with early pricing

The critical insight: users check weather forecasts before making booking decisions. By proactively adjusting prices when rain probability exceeds threshold, we capture 78% of at-risk outdoor bookings as indoor facility transfers—revenue that would otherwise be lost to cancellation.

3

Module 2: Dynamic Pricing Engine

We developed a LightGBM-based pricing model that predicts optimal price points using 19 engineered features. The model is constrained to ±60% from base priceto maintain perceived fairness while maximizing revenue capture.

LightGBM Architecture

Gradient Boosting Decision Tree

Model Parameters

  • Trees500
  • Max depth8
  • Learning rate0.05
  • Feature fraction0.8

Training Data

  • Samples2.4M
  • Time range18 months
  • Facilities340+
  • CV folds5

3.1 Feature Engineering (19 Features)

Temporal (5)
  • day_of_week
  • hour_of_day
  • is_holiday
  • time_to_slot (hours)
  • is_weekend
Demand (7)
  • current_occupancy
  • hist_demand_7d
  • hist_demand_30d
  • booking_velocity
  • cancellation_rate
  • same_hour_hist
  • same_dow_hist
External (7)
  • rain_probability
  • temperature
  • wind_speed
  • is_indoor
  • facility_rating
  • nearby_events
  • competitor_price

Price Constraint: ±60% from Base

Unlike airline pricing (which can vary 400%+), we constrain price adjustments to maintain user trust and perceived fairness. Research shows sports facility users are more price-sensitive to variance than absolute price.

MinimumBaseMaximum
₩12,000₩30,000₩48,000
4

Module 3: Hybrid Recommendation System

We developed a hybrid recommendation engine combining collaborative filtering (CF) with content-based (CB) approaches. The system predicts facility preferences using matrix factorization and adjusts recommendations based on real-time weather conditions.

CF + CB Hybrid Architecture

Matrix Factorization with Cosine Similarity

Collaborative Filtering

User-user similarity based on booking history and facility ratings. Identifies "users like you" to surface facilities they enjoyed.

sim(u, v) = cos(R_u, R_v)

Content-Based

Facility attributes (surface type, amenities, location, price tier) matched against learned user preference vectors.

score = w_cf × CF + w_cb × CB

Weather-Aware Adjustment

When rain probability exceeds 50%, the system automatically boosts indoor facility scores by 40% and demotes outdoor facilities proportionally. Users with high outdoor preference receive proactive indoor alternatives.

Output Format

The system returns Top 5 facility recommendations with predicted match scores:

{
  "recommendations": [
    { "facility_id": "F001", "score": 0.94, "reason": "similar_users" },
    { "facility_id": "F023", "score": 0.89, "reason": "weather_adjusted" },
    { "facility_id": "F017", "score": 0.87, "reason": "content_match" },
    ...
  ],
  "indoor_preference": 0.73,
  "weather_adjusted": true
}
5

Results & TIPS Certification

TIPS Certification Achieved

Tech Incubator Program for Startup Korea (기술창업기업 TIPS)

98%
Prediction Accuracy
+23%
Revenue Uplift
-41%
Empty Slot Rate
16
Cities Deployed

System validated across 340+ facilities over 18-month deployment period. A/B testing showed statistically significant (p < 0.01) revenue improvements in dynamic pricing cohort vs. static pricing control.

The TIPS certification validates that sports facilities can successfully adopt airline-style yield management. The ±60% price constraint maintains perceived fairness while capturing significant revenue optimization opportunities. Key success factors include:

  • Transparent pricing communication: Users understand why prices vary
  • Weather-adjusted recommendations: Genuinely helpful, not manipulative
  • Consistent value proposition: "Fair deal" perception maintained
6

Technical Implementation

Data Layer

  • • Supabase PostgreSQL
  • • KMA ASOS API (실시간)
  • • Redis caching layer
  • • Event-driven ETL

ML Pipeline

  • • Python FastAPI
  • • LightGBM models
  • • scikit-learn preprocessing
  • • MLflow experiment tracking

Deployment

  • • Supabase Edge Functions
  • • Next.js 14 frontend
  • • Flutter mobile app
  • • Vercel + AWS hybrid

References

[1] Talluri, K. & Van Ryzin, G. (2004). The Theory and Practice of Revenue Management. Springer Science.

[2] Ke, G. et al. (2017). LightGBM: A Highly Efficient Gradient Boosting Decision Tree. NeurIPS.

[3] Koren, Y. et al. (2009). Matrix Factorization Techniques for Recommender Systems. IEEE Computer.

[4] Smith, B. et al. (1992). Yield Management at American Airlines. Interfaces 22(1):8-31.